Computational validation demonstrates superior defect segmentation across diverse textile textures, indicating the effectiveness of texture-aware attention mechanisms.
Key Points
To develop a Swin Transformer framework capable of accurately segmenting multi-scale fabric surface defects across complex textures and low-contrast conditions in industrial visual inspection.
Designed a framework incorporating a Texture Enhancement Module (TEM), a Defect-Guided Attention (DGA) mechanism, and a Hybrid Multi-scale Context (HMC) module coupled with a UPerNet decoder.
Trained and evaluated the model on the ZJU-Leaper textile defect benchmark dataset across four representative fabric groups.
Consistently outperformed state-of-the-art segmentation methods across all four evaluated fabric groups.
Achieved superior pixel-level defect delineation under challenging scenarios featuring weak visual contrast and intricate background textures.